Waterr Research
AI research that puts transparency in every interaction at the frontier.
Realtime voice models trade reasoning depth for latency. Our research closes that gap — and everything on this page runs in production meetings today, or says plainly that it doesn’t yet.
Flagship · Harness
Monologue
Shipping in every meetingThe agent talks to itself while it listens.
A reasoning harness for realtime voice models: while the participant speaks, a concurrent planner aligns on facts, commitments, and where the conversation should go next — then briefs the voice model in the gap between turns. Added intelligence, no added latency. It’s what keeps long conversations meaningful past turn six.
Measured on Audio MultiChallenge
+25%
relative improvement over the bare voice model on Scale AI’s AudioMC benchmark (+9.6 pts paired APR, n=124 conversations).
Models
ORI — the speech stack
One family, two surfaces: bring your own agent, or build on our realtime model. Every meeting Waterr runs is a live deployment.
Voice for your agents
ORI Utter
AvailableBrings voice to your text agents with Waterr’s speech stack. We handle audio orchestration, deployment, and observability — you keep your agent’s reasoning.
Realtime interaction model
ORI Realtime
Research previewA realtime model for natural conversation on a single base architecture — built to cut latency without giving up reasoning depth. Exposed to developers over WebSocket.
Model card & docs →Deployment
Runs where your conversations do.
Cloud
Managed API endpoints with in-region processing, high availability, and the uptime production systems demand.
On-prem
Run the stack inside your VPC or on your own hardware. Your data never leaves your perimeter.
On-device
A realtime stack for on-device, end-to-end video calls — proven in the lab, opening to early partners.
The research is in the room.
Every claim on this page is one meeting away from proof. Talk to an agent, or read the notes from the lab.